Stepping off the hedonic treadmill: individual differences in response to major life events

Journal of Individual Differences 32(3): 144–152 (2011), DOI 10.1027/1614-0001/a000047. Read as the publisher PDF posted on the Loss, Trauma and Emotion Lab site at Teachers College, Columbia. Pace University, Teachers College and Paris School of Economics. Provenance: papers/carry_on/mancini2011_hedonic_treadmill.provenance.json.

Read it together with summaries/carry_on/infurna2016_resilience.md. Infurna & Luthar (2016) re-ran exactly these models on the same panel. With the constraints used here relaxed, the "resilient" majorities shrank (e.g. divorce 36% resilient instead of 71.8%).

What was read

All 550 lines of pdftotext -layout output (9 pages): abstract, all sections, Tables 1–3, figure captions, acknowledgements and references. Figures 1–3 (the trajectory plots) are images and were not inspected; class shapes are known only from the text and the growth parameters. The fit indices for the one- to five-class models are "available from first author" and are not in the paper.

Question

Is the response to major life events one average trajectory with noise, or several distinct trajectories? Latent growth mixture models (LGMM) are used to find them.

Method

• Data. German Socio-Economic Panel (SOEP), waves 1984–2003. The samples are people who reported widowhood, divorce or marriage between 1985 and 2003, aged 75 or under at the event.
  • Bereaved: 464, 76% women, mean age 60.
  • Divorced: 629, mean age 40.
  • Married: 1,739, mean age 28.
• Window. Nine annual waves per event: four before, the event year, four after.
• Outcome. Single-item life satisfaction, 0–10.
• Covariates. Age, sex, education, household income (mean and pre-to-post change) and health dysfunction (one item averaged over nine waves, α .93). Only covariates that improved fit were kept.
• Model.
  • Specification: freely estimated time scores for bereavement and divorce; linear plus quadratic for marriage.
  • Classes: one to five were compared on BIC, SSBIC, AIC, entropy and LRT/BLRT, and on theory and interpretability.
  • Fixed variances: for bereavement and divorce, slope variance was fixed at zero, with intercept variance free across classes. For marriage the quadratic variance was fixed at zero.

Results

• Bereavement: four classes (Table 1; the intercept is the level, the slope is change over the window).
  • Resilient: 58.7%, stable around 7.7.
  • Acute-recovery: 21.3%, high before, a sharp dip, then a gradual return.
  • Chronic low: 14.6%, low before (5.2), a dip, then back.
  • Improved: 5.4%, a sharp rise at the loss, then a gradual decline.
  • Covariates: the resilient were older, healthier and lost less income. The improved gained income.
• Divorce: three classes (Table 2).
  • Resilient: 71.8%, flat at 7.0.
  • Moderate-decreasing: 19.1%, declining from well before the divorce.
  • Low-increasing: 9.1%, low before (3.8), then rising sharply (slope 3.03).
  • Covariates: health dysfunction separates both other classes from the resilient; education only marginally (p = .10).
• Marriage: four classes (Table 3).
  • High-stable: 79.6%.
  • Decreasing-increasing: 9.1%, falling before the marriage and rising after.
  • Decreasing: 6.0%, falling sharply after the marriage.
  • Increasing: 5.2%, rising before the marriage and staying up.
  • Covariates: health dysfunction separates every class from high-stable; lower income separates two of them.
• Sex predicted no class membership.
• Authors' reading.
  • The modal response to each event is no change.
  • A treadmill pattern (react, then return) appears only in a minority.
  • Some people improve after bereavement or divorce, and some decline after marriage.
  • "The impact of life events may depend largely on individual differences and on life circumstances."

Limits

• Constraints drive the class shares. Fixing the slope variance at zero within classes forces everyone in a class onto one trajectory shape, so heterogeneity must be absorbed by extra classes or put into the stable one. Infurna & Luthar (2016) showed the resilient share falls sharply when this is relaxed.
• The number of classes was chosen partly on theory and interpretability, with fit indices not shown.
• Single-item life satisfaction, measured annually; no personality or appraisal measures.
• Covariates are coarse and partly outcomes themselves (income change, averaged health).
• Inconsistencies found:
  • Throughout, p-values are printed as "p > .001" where "p < .001" is meant: 11 occurrences, e.g. χ²(9, N = 455) = 100.88, "p > .001".
  • The Ns in the covariate tests (455, 625, 1,733) are below the sample sizes (464, 629, 1,739), without explanation.
  • Some class labels do not match the growth parameters.
    • The bereavement "improved" class slope is 1.64, with a CI including zero (−.13 to 3.41).
    • "Chronic low" is described as dipping and returning, but its slope is −.44 (n.s.) under freely estimated time scores.
    • The marriage "decreasing" class has a positive linear term (+0.70) and a negative quadratic (−0.13). It is an inverted U, a rise then a fall, as the "increasing" class also is (+1.50, −0.14).
  • Arithmetic checked: class shares sum to 100.0, 100.0 and 99.9%.

What it means for Kurisutina

• Question 1: reactions to the same event diverge in direction, not only in size (verified from the paper).
  • After widowhood, a minority improves. After divorce, 9.1% rise from a low start. After marriage, some decline.
  • A replica that applies the average reaction to an event type will be wrong in sign for these people. The slot has to say which kind of person this is for this kind of event.
• Much of the divergence is visible before the event.
  • Chronic-low bereaved and low-increasing divorced start low. Moderate-decreasing divorced and decreasing-increasing married are already moving before the event.
  • This matches Infurna (resilient classes start higher and are less volatile) and Luhmann 2014 (change starts before the event).
  • For the replica, the pre-event trajectory in the slot, not just the pre-event level, is informative (inferred). The GSS pilot gives only one earlier wave, so it cannot carry this. LISS monthly data could.
• Circumstances predict class, personality was not tested. Health and income change (resources) separated the classes. That fits Sliwinski's finding that current load moves reactivity. A replica needs the person's circumstances, not only their dispositions (inferred).
• Class shares are model artefacts to a degree. Do not use "58.7% resilient" or similar as a prior for a replica. The continuous-heterogeneity reading (Infurna) is the safer default for LISS T1: estimate each person's reaction as a continuous quantity, as T1 already does.

Cross-references

• summaries/carry_on/infurna2016_resilience.md: the re-analysis of these models with relaxed constraints.
• summaries/carry_on/luhmann2014_its_about_time.md: pre-event change and reversible change.
• summaries/carry_on/sliwinski2009_stress_bursts.md: current load and reactivity.
• summaries/carry_on/luhmann2012_adaptation.md: average reactions to the same events.
• docs/research/liss_q1_design.md: T1 reaction estimates.